feat(klausur): Handschrift entfernen + Klausur-HTR implementiert

Feature 1: Handschrift entfernen via OCR-Pipeline Session
- services/handwriting_detection.py: _detect_pencil() + target_ink Parameter
  ("all" | "colored" | "pencil") für gezielte Tinten-Erkennung
- ocr_pipeline_session_store.py: clean_png + handwriting_removal_meta Spalten
  (idempotentes ALTER TABLE in init_ocr_pipeline_tables)
- ocr_pipeline_api.py: POST /sessions/{id}/remove-handwriting Endpoint
  + "clean" zu valid_types für Image-Serving hinzugefügt

Feature 2: Klausur-HTR (Hochwertige Handschriftenerkennung)
- handwriting_htr_api.py: Neuer Router /api/v1/htr/recognize + /recognize-session
  Primary: qwen2.5vl:32b via Ollama, Fallback: trocr-large-handwritten
- services/trocr_service.py: size Parameter (base | large) für get_trocr_model()
  + run_trocr_ocr() - unterstützt jetzt trocr-large-handwritten
- main.py: HTR Router registriert

Config:
- docker-compose.yml: OLLAMA_HTR_MODEL, HTR_FALLBACK_MODEL
- .env.example: HTR Env-Vars dokumentiert

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

klausur-service/backend/main.py

@@ -44,6 +44,10 @@ except ImportError:
 from vocab_worksheet_api import router as vocab_router, set_db_pool as set_vocab_db_pool, _init_vocab_table, _load_all_sessions, DATABASE_URL as VOCAB_DATABASE_URL
 from ocr_pipeline_api import router as ocr_pipeline_router
 from ocr_pipeline_session_store import init_ocr_pipeline_tables
+try:
+    from handwriting_htr_api import router as htr_router
+except ImportError:
+    htr_router = None
 try:
     from dsfa_rag_api import router as dsfa_rag_router, set_db_pool as set_dsfa_db_pool
     from dsfa_corpus_ingestion import DSFAQdrantService, DATABASE_URL as DSFA_DATABASE_URL
@@ -113,6 +117,19 @@ async def lifespan(app: FastAPI):
     # Ensure EH upload directory exists
     os.makedirs(EH_UPLOAD_DIR, exist_ok=True)
 
+    # Preload LightOnOCR model if OCR_ENGINE=lighton (avoids cold-start on first request)
+    ocr_engine_env = os.getenv("OCR_ENGINE", "auto")
+    if ocr_engine_env == "lighton":
+        try:
+            import asyncio
+            from services.lighton_ocr_service import get_lighton_model
+            loop = asyncio.get_event_loop()
+            print("Preloading LightOnOCR-2-1B at startup (OCR_ENGINE=lighton)...")
+            await loop.run_in_executor(None, get_lighton_model)
+            print("LightOnOCR-2-1B preloaded")
+        except Exception as e:
+            print(f"Warning: LightOnOCR preload failed: {e}")
+
     yield
 
     print("Klausur-Service shutting down...")
@@ -160,6 +177,8 @@ if trocr_router:
     app.include_router(trocr_router)      # TrOCR Handwriting OCR
 app.include_router(vocab_router)      # Vocabulary Worksheet Generator
 app.include_router(ocr_pipeline_router)  # OCR Pipeline (step-by-step)
+if htr_router:
+    app.include_router(htr_router)            # Handwriting HTR (Klausur)
 if dsfa_rag_router:
     app.include_router(dsfa_rag_router)   # DSFA RAG Corpus Search